LangGraph Platform

If you are evaluating something called "LangGraph Platform", you are evaluating LangSmith Deployment. LangChain's own FAQ on the page states it plainly: "Yes, LangGraph Platform has been renamed to LangSmith Deployment as of October 2025." The marketing URL langchain.com/langgraph-platform now returns a 301 to langchain.com/langsmith/deployment, and the documentation path docs.langchain.com/langgraph-platform/deployment-options now serves a page titled "Set up LangSmith".

The practical consequence is that the product is no longer sold as a standalone thing. Deployment is one service inside LangSmith, alongside observability, evaluation, Fleet, Engine and Sandboxes, and it is billed on the same LangSmith plan and the same metering units. If your evaluation notes, procurement paperwork or an internal comparison doc say "LangGraph Platform", they are describing a line item that no longer exists under that name. Third-party pricing articles are especially prone to this — several still quote a "$35/month LangGraph Platform" figure that does not correspond to anything on the current pricing page.

Separately, and more easily confused: LangGraph itself is the open-source orchestration framework, and it is not this product. LangChain's FAQ draws the line the same way — LangGraph is the framework for building stateful agent workflows, LangSmith Deployment is the managed service for running them.

The core argument for this product is that agents in production need a specific set of unglamorous infrastructure, and it ships that set rather than making you assemble it. The vendor frames it as durable execution, streaming, state and task management, and the documented feature set backs that up: assistants for configuration, threads for state, and runs for workloads, with human-in-the-loop approvals, background tasks and conversation threading exposed as APIs.

Two things are worth singling out because they are not obvious from the name.

First, it is framework-agnostic, and specifically so. The docs state that LangSmith Deployment can host agents built with the Claude Agent SDK, Strands, CrewAI, AutoGen and other frameworks via the Functional API or a deployments-wrap-sdk wrapper. You are not required to adopt LangGraph to use the runtime, which materially widens who this is for and undercuts the usual assumption that it is a LangChain-ecosystem lock-in play at the framework layer.

Second, the underlying open-source framework is unusually healthy, which matters for anyone betting on the ecosystem's continuity. The langgraph repository is MIT-licensed, carries 39,694 stars, and was last pushed on 14 August 2026 — the day before this listing was written. It is well past 1.0, with release 1.2.11 published on 11 August 2026. This is not a project coasting on an old release.

On the operational side, Dedicated deployments autoscale across replicas on CPU utilisation, memory utilisation and pending run count, with queue workers scaling on pending runs while API servers scale on CPU and memory — so read traffic does not slow run submission. Size changes roll out as a new revision with no downtime.

Pricing is metered in two abstract units, which makes the sticker price hard to read. The vendor defines them: 1 LangChain Compute Unit (LCU) = $1.50, and 1 LangChain Storage Unit (LSU) = $1.00. Deployment metering rates are published as 0.045 LCU per vCPU-hour and 0.006 LCU per GiB-hour for runtime, and 0.177 LSU per vCPU-hour and 0.025 LSU per GiB-hour for the database.

The docs publish the resources for each deployment size separately. Combining the two — which the vendor does not do on a single page — gives the monthly floor for an always-on Dedicated deployment at 730 hours, at minimum replica count:

• Dedicated Small (3 vCPU / 6 GiB runtime, 1 vCPU / 4 GiB database): about $389 per month
• Dedicated Medium (5 vCPU / 12 GiB runtime, 2 vCPU / 8 GiB database): about $730 per month
• Dedicated Large (10 vCPU / 24 GiB runtime, 4 vCPU / 16 GiB database): about $1,459 per month

These are floors, not estimates. Dedicated deployments are always-on and consume compute continuously, and autoscaling adds replicas on top.

The detail worth knowing: on Dedicated Small, the database accounts for about $202 of the roughly $389, or over half the bill. Database compute is metered at 0.177 LSU/vCPU-hour, which is roughly 2.6x the runtime compute rate per vCPU-hour once you convert both to dollars. The persistent Postgres that gives you durable state is the expensive half of the product, and it does not scale down.

Serverless is the cheaper path and has no dedicated database line. A Serverless Medium (2 vCPU / 5 GiB) running continuously computes to about $131 per month, less if it idles, because Serverless scales to zero after inactivity. Two caveats: scale-to-zero is labelled beta in the billing docs, and compute is billed while resources are provisioned, so charges stop only once it has actually scaled down, not at the moment traffic stops. Serverless also runs on shared multi-tenant infrastructure and the first request after scale-down is slower.

On top of deployment: seats are $0 on Developer (1 seat maximum) and $39 per seat per month on Plus. Plus includes one free Serverless (Small) deployment, which is a real starting allowance. Traces are billed separately — 5k base traces/month included on Developer, 10k on Plus, then 0.005 LSU per additional trace, which is $0.005 per trace or $5 per 1,000. Note that base traces are retained for 14 days; the 400-day extended retention costs more.

This is the distinction most likely to catch out a team that assumes "LangGraph is open source" settles the question.

The framework, langgraph, is MIT — confirmed both on the GitHub repository and in the PyPI metadata. You can build on it, fork it and run it anywhere with no strings.

The server is a different matter. The langgraph-api package on PyPI — the local development server you get from langgraph dev — is licensed under Elastic License 2.0, not MIT. ELv2 is a source-available licence that restricts providing the software to third parties as a hosted or managed service and prohibits circumventing licence-key functionality. Worth being precise about what this package is: PyPI describes it as the server for rapid development and testing, backed by a predominantly in-memory store persisted to local disk. It is the dev server, not the production runtime.

The production runtime is licensed commercially. Running the standalone Agent Server yourself requires a LANGGRAPH_CLOUD_LICENSE_KEY, which the docs describe as a LangSmith licence key authenticated at server startup.

So the accurate summary is a three-tier split: the orchestration framework is genuinely open source, the development server is source-available under ELv2, and the production runtime that makes this a platform requires a paid licence. None of that is hidden, but none of it is on the marketing page either.

Self-hosted and hybrid deployment options are listed as Enterprise plan features on the pricing page. There is no self-serve path to running this in your own VPC — Developer and Plus are cloud-only for deployment.

If you do get an Enterprise contract, self-hosted is not fully disconnected by default. The docs are explicit: unless you are running in offline mode, self-hosted LangSmith requires egress to https://beacon.langchain.com for billing telemetry, covering licence verification and subscription/usage reporting. That billing telemetry is marked required and "cannot be disabled" — the documented route to running with no egress is to contact your account team about an offline (air-gapped) licence. Operational telemetry, which exports metrics, traces and error log messages, is separately disableable; disabling it does not stop the billing telemetry.

One functional limit if you go air-gapped: Engine cannot run offline, because it depends on LangSmith Intelligence, a LangChain-managed service outside your environment. The docs state every other LangSmith feature continues to work offline.

For most teams none of this is disqualifying — licence check-ins are standard for commercial self-hosted software. It matters if you are in a regulated or genuinely air-gapped environment, because it turns a technical question into a contract negotiation with your account team.

The LCU/LSU model described above is the current pricing. Organisations on the previous pricing are in a transition window: the docs state they can continue creating the older Development and Production deployment types until 1 October 2026, and note that those older types do not include scale to zero.

If you are an existing customer, this is the number to check before it moves — the older deployment types were billed differently, and the transition deadline is roughly six weeks out from the date of this listing. If you are a new buyer, you are on the LCU/LSU model and the older figures circulating in third-party comparisons do not apply to you.

LangChain announced a $125M raise at a $1.25B valuation, led by IVP alongside existing investors Sequoia, Benchmark and Amplify, with CapitalG and Sapphire Ventures joining. That is from the company's own announcement, and it is the round that also carried the LangChain and LangGraph 1.0 releases. Press coverage dates the round to October 2025 and reports a broader investor list including ServiceNow, Workday, Cisco, Datadog and Databricks; the company's own post names only the investors listed above, so treat the wider list as reported rather than confirmed. Continuity risk is low by startup standards — this is a well-capitalised company, not a one-person project.

On adoption, the deployment page carries a large customer logo wall including Klarna, LinkedIn, Uber, Cisco, Nvidia, Workday, Coinbase, Rakuten, ServiceNow, Bristol Myers Squibb, Harvey and Elastic. These are vendor claims presented without case studies on that page, and a logo does not establish that the company uses this specific deployment product rather than the open-source framework or LangSmith tracing. Independent, verifiable adoption evidence specific to the deployment product is not established.

The strongest independent signal available is the framework's own activity, covered above: 39,694 stars and near-daily releases. Independent review-site ratings for the deployment product could not be retrieved, and no reliable third-party review data was found — most search results for pricing and comparisons are published by direct competitors or affiliate content farms, which is worth knowing if you go looking yourself.

It suits you if you are running agents that are long-lived, interruptible or human-approved — the durable execution and thread/state model is the real product, and rebuilding it is a genuine engineering project. It suits you if you want managed hosting without adopting LangGraph as your framework, since the runtime accepts other frameworks. And it suits a team already paying for LangSmith observability, because deployment now shares the plan and the meter.

Weigh it more carefully if your agent is a simple request/response service with no durable state, where a container on your existing platform costs a fraction of $389/month. Weigh it if you need self-hosting without an Enterprise contract, which is not available. And weigh it if you run many separate always-on deployments, because each Dedicated deployment carries its own database and that database is over half the per-deployment cost — the model rewards consolidating agents behind fewer deployments, and it does not reward a deployment per microservice.

The free Serverless Small deployment on the $39 Plus plan is a genuinely usable way to test all of this before committing, and it is the sensible first step rather than modelling the cost in a spreadsheet.